Enterprise buyers are pushing back on expensive frontier APIs, and early usage numbers for copilots are nowhere near justifying the current AI infra spend. Governments are moving from regulating AI to owning and weaponizing it, while data centers run into carbon, grid, and community limits and China builds a cheaper parallel stack.
The easy AI trade is over; the next phase is about whose economics and politics actually hold up under stress.
Key Events
/OpenAI is negotiating to give the US government a 5% equity stake worth about $42.6B at an $852B valuation.
/Microsoft cut 4,800 jobs, including about half its Xbox division, and is replacing OpenAI/Anthropic models in Office apps with its own AI to lower costs.
/Spain officially blacklisted Palantir from both public and private sectors over national security concerns.
/The EU’s top court upheld a record €4.1B (~$4.7B) antitrust fine against Google over Android dominance.
/SpaceX deorbited 260 Starlink satellites in six months and asked the FCC to approve a third‑generation constellation of 100,000 satellites.
Report
AI is shifting from a free‑money story to a unit‑economics story just as governments stop acting like referees and start joining the game as owners and censors.
OpenAI is floating a 5% equity stake to the US, Microsoft is ripping pricey frontier APIs out of Office, Spain is exiling Palantir, and communities are suing to halt AI data centers.
the frontier-api backlash and margin squeeze
Palantir CEO Alex Karp went on TV to call closed frontier models "completely oversold" and "intellectual property extraction dressed up as a subscription," pointing directly at token pricing.
He says enterprises increasingly want to "own the means of production" instead of streaming operational data to OpenAI- or Anthropic-style APIs.
Microsoft is already swapping OpenAI and Anthropic models out of Excel and Outlook in favor of its own models explicitly to cut external AI costs.
Microsoft 365 Copilot adoption is under 4.5%. Only about 1% of Copilot users engage weekly despite years of promotion and price hikes, which is not the usage curve you build $11T compute narratives on.
the ai bubble test in real time
The US Treasury’s internal report now explicitly warns about the risk of an AI bubble. Wall Street analysis is framing the AI compute boom as an $11T wave that could leave markets holding a $7T debt pile if returns disappoint.
The "Magnificent 7" have already shed about $2.3T in value on AI-spend worries, while the BIS is flagging systemic risks in venture and AI exposure.
A Marxist‑Leninist group has managed to delay or block $23.6B of US AI infra projects, including multiple data-center moratoria, so not all the capex in pitch decks is turning into concrete.
Spain’s solar sector is now so oversupplied and cheap that some investors are looking to exit projects, a live example of how a hot infra story can flip once capacity outruns demand.
regulators as shareholders, censors, and threats
OpenAI is negotiating to hand the US government a 5% equity stake worth around $42.6B at an $852B valuation to buy cooperation and reduce political risk.
On the culture‑war front, Peter Thiel is attacking the Pope as a "Chinese communist agent" for backing AI regulation, while the US is labeling some anti‑AI activism as "anti‑tech extremism"—AI policy is now partisan terrain, not a neutral standards process.
In Europe, lawmakers keep pushing Chat Control–style scanning of private messages and are shipping AI Act tooling like an open‑source compliance scanner, while courts uphold multibillion‑euro antitrust fines against Google over Android.
Japan’s top court has ruled that AI cannot be listed as an inventor on patents, drawing a bright line against machine rights in IP. Spain has gone further by blacklisting Palantir from both public and private sectors on national‑security grounds even as Palantir deepens classified AI deployments with Nvidia inside US government systems.
data centers hit the community and carbon wall
Amazon’s carbon emissions jumped 16% in 2025, driven largely by a record data center buildout for cloud and AI. Google expects its AI expansion to drive a 37% increase in electricity use in 2025, meaning AI shows up visibly on utility load curves.
AI data centers are already drawing lawsuits and bans: Wisconsin residents filed a class action over noise from Microsoft’s $7.3B facility, and San Marcos, Texas, became the first US city to outright ban data centers.
Meta had water discharges from a data center suspended after contaminating a city’s reclaimed water with bacteria, and investigative work says AI data centers use far more water than operators publicly disclose.
Local governments are asking employees to conserve power as AI demand drives up electricity prices, even as DHS and FBI recast organized opposition to AI and data centers as a potential form of extremism.
china, cheap silicon, and a parallel ai stack
China plans to phase out foreign‑made chips by 2027 while adding 543 GW of new grid capacity in a year, including 434 GW of renewables, to power its own compute.
Chinese AI models are now seen as "good enough" for many workloads at much lower prices, with at least one inexpensive model catching up with OpenAI and Anthropic on core benchmarks.
Qwen3‑235B fine‑tuned on Tinker hit 84.7% accuracy on document filtering, beating the best frontier models while being significantly cheaper to run.
Beijing is openly considering restricting overseas access to its top models and criminalizing leakage of proprietary AI technology as a national‑security crime, pointing toward a closed, sovereign stack.
At the same time, cheap Chinese RAM and chips are poised to ease the global memory price crunch even as China pushes to eliminate reliance on foreign silicon, reshaping both cost curves and dependency maps.
What This Means
The easy AI trade—pay whatever it takes for closed models and hyperscale infra and assume demand will catch up—is colliding with buyer skepticism, state control, and physical limits at the same time. The live decision is whether to stay levered to that high‑burn, high‑dependence path or pivot toward cheaper, more sovereign, and more resource‑efficient stacks before margin and regulatory shocks fully price in.
On Watch
/Cloudflare’s September 15 move to block most AI agents and training bots by default on roughly a fifth of the web could quietly choke off cheap training data and agent reach.
/SpaceX’s proposed 100,000‑satellite Starlink Gen3 network, alongside Amazon’s planned satellite internet service, could turn LEO broadband into a strategic gatekeeper for AI services in emerging markets.
/Japan’s top‑court ban on AI inventorship, combined with a US judge canceling a trial over AI misuse, hints at a hard ceiling on where courts will accept AI as an autonomous actor.
Interesting
/Meta employees consumed 73.7 trillion AI tokens in one month, costing around $221 million.
/Over 95% of the Grace-Blackwell GPUs have not yet been deployed, despite shipping since December 2024.
/Bridgewater's hedge fund found that no frontier AI model met the 80% accuracy threshold needed for investor trust, raising concerns about AI reliability.
/46% of doctoral-level graduates in US science and engineering are foreign-born, highlighting the reliance on international talent in critical fields.
/The U.S. Commerce Department's move to restrict access to frontier models from companies like Anthropic and OpenAI parallels China's considerations to limit foreign access to its AI technologies.
We processed 10,000+ comments and posts to generate this report.
AI-generated content. Verify critical information independently.
/OpenAI is negotiating to give the US government a 5% equity stake worth about $42.6B at an $852B valuation.
/Microsoft cut 4,800 jobs, including about half its Xbox division, and is replacing OpenAI/Anthropic models in Office apps with its own AI to lower costs.
/Spain officially blacklisted Palantir from both public and private sectors over national security concerns.
/The EU’s top court upheld a record €4.1B (~$4.7B) antitrust fine against Google over Android dominance.
/SpaceX deorbited 260 Starlink satellites in six months and asked the FCC to approve a third‑generation constellation of 100,000 satellites.
On Watch
/Cloudflare’s September 15 move to block most AI agents and training bots by default on roughly a fifth of the web could quietly choke off cheap training data and agent reach.
/SpaceX’s proposed 100,000‑satellite Starlink Gen3 network, alongside Amazon’s planned satellite internet service, could turn LEO broadband into a strategic gatekeeper for AI services in emerging markets.
/Japan’s top‑court ban on AI inventorship, combined with a US judge canceling a trial over AI misuse, hints at a hard ceiling on where courts will accept AI as an autonomous actor.
Interesting
/Meta employees consumed 73.7 trillion AI tokens in one month, costing around $221 million.
/Over 95% of the Grace-Blackwell GPUs have not yet been deployed, despite shipping since December 2024.
/Bridgewater's hedge fund found that no frontier AI model met the 80% accuracy threshold needed for investor trust, raising concerns about AI reliability.
/46% of doctoral-level graduates in US science and engineering are foreign-born, highlighting the reliance on international talent in critical fields.
/The U.S. Commerce Department's move to restrict access to frontier models from companies like Anthropic and OpenAI parallels China's considerations to limit foreign access to its AI technologies.